Tackling the “curse of dimensionality” of radial basis functional neural networks using a genetic algorithm

Abstract

Radial Basis Function (RBF) neural networks offer the possibility of faster gradient-based learning of neuron weights compared with Multi-Layer Perceptron (MLP) networks. This apparent advantage of RBF networks is bought at the expense of requiring a large number of hidden layer nodes, particularly in high dimensional spaces (the “curse of dimensionality”). This paper proposes a representation and associated genetic operators which are capable of evolving RBF networks with relatively small numbers of hidden layer nodes and good generalisation properties. The genetic operators employed also overcome the “competing conventions” problem, for RBF networks at least, which has been a reported stumbling block in the application of crossover operators in evolutionary learning of directly encoded neural network architectures.